In [1]:
%matplotlib inline
%config InlineBackend.figure_format='retina'

from IPython.display import display, display_markdown

import warnings
warnings.simplefilter(action='ignore', category=FutureWarning)

import subprocess as sp
import numpy as np
import pandas as pd
import seaborn as sns
import arviz as az
import bambi
from copy import deepcopy
from scipy.stats import pearsonr, spearmanr

import matplotlib.pyplot as plt
plt.rcParams['figure.figsize'] = [10, 8]
plt.rcParams['figure.dpi'] = 150

from subs2vec.utensils import log_timer
from subs2vec.vecs import Vectors
from subs2vec.neighbors import compute_nn

import logging
logging.getLogger().setLevel(logging.WARNING)

def display_md(md, **kwargs):
    return display_markdown(md, raw=True, **kwargs)

def convert_notebook(title, output='html'):
    convert = sp.run(f'jupyter nbconvert {title}.ipynb --to {output} --output {title}.{output}'.split(' '))
    if convert.returncode == 0:
        display_md(f'Jupyter notebook `{title}` converted successfully.')
    else:
        display_md(f'Error: encountered problem converting Jupyter notebook `{title}`')

def norm(x):
    return x / np.linalg.norm(x, 2)

def standardize(x):
    return (x - x.mean()) / x.std()

sns.set(style='whitegrid')
pd.options.mode.chained_assignment = None

Loading and preparing data¶

In [2]:
df_joint = pd.read_csv('data/data_plus_predictors.tsv', sep='\t')
df_joint = df_joint[df_joint['dimension'] != 'high-low']
display(df_joint)
index group dimension pp_id color rating experiment self_vs_other art fiction ... sighted group_eff group_z original replication_1 replication_2 other self self_vs_other_eff self_vs_other_z
0 0 sighted cold-hot sighted_1 white 1 original self NaN NaN ... 1 1.0 0.211241 1 0 0 0 1 -1.0 -0.891882
1 1 sighted ripe-unripe sighted_1 white 7 original self NaN NaN ... 1 1.0 0.211241 1 0 0 0 1 -1.0 -0.891882
2 2 sighted new-old sighted_1 white 1 original self NaN NaN ... 1 1.0 0.211241 1 0 0 0 1 -1.0 -0.891882
3 3 sighted submissive-aggressive sighted_1 white 1 original self NaN NaN ... 1 1.0 0.211241 1 0 0 0 1 -1.0 -0.891882
4 4 sighted selfless-jealous sighted_1 white 1 original self NaN NaN ... 1 1.0 0.211241 1 0 0 0 1 -1.0 -0.891882
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
42974 28506 sighted soft-hard sighted_68129 red 5 replication_2 other 11.0 0.0 ... 1 1.0 0.211241 0 0 1 1 0 1.0 1.121199
42975 28507 sighted light-heavy sighted_68129 red 5 replication_2 other 11.0 0.0 ... 1 1.0 0.211241 0 0 1 1 0 1.0 1.121199
42976 28508 sighted relaxed-tense sighted_68129 red 5 replication_2 other 11.0 0.0 ... 1 1.0 0.211241 0 0 1 1 0 1.0 1.121199
42977 28509 sighted alive-dead sighted_68129 red 6 replication_2 other 11.0 0.0 ... 1 1.0 0.211241 0 0 1 1 0 1.0 1.121199
42978 28510 sighted fast-slow sighted_68129 red 3 replication_2 other 11.0 0.0 ... 1 1.0 0.211241 0 0 1 1 0 1.0 1.121199

40866 rows × 74 columns

How similar are the ratings in our replication data to the original Saysani sighted sample?¶

In [3]:
replication = df_joint[df_joint['experiment'] == 'replication_1'].groupby(['dimension', 'color']).mean()
original = df_joint[(df_joint['experiment'] == 'original') & (df_joint['group'] == 'sighted')].groupby(['dimension', 'color']).mean()
r = pearsonr(replication['rating'], original['rating'])
print(f'r(original, replication) = {r[0]:.2f}, p = {r[1]:.3f}')
r = spearmanr(replication['rating'], original['rating'])
print(f'rank r(original, replication) = {r[0]:.2f}, p = {r[1]:.3f}')

r(original, replication) = 0.89, p = 0.000
rank r(original, replication) = 0.87, p = 0.000

How similar are blind and sighted participants' ratings?¶

In [4]:
blind = df_joint[df_joint['group'] == 'blind'].groupby(['dimension', 'color']).mean()
sighted = df_joint[((df_joint['experiment'] == 'original') | (df_joint['experiment'] == 'replication_1'))
                   & (df_joint['group'] == 'sighted')].groupby(['dimension', 'color']).mean()
r = pearsonr(blind['rating'], sighted['rating'])
print(f'r(blind, sighted) = {r[0]:.2f}, p = {r[1]:.3f}')
r = spearmanr(blind['rating'], sighted['rating'])
print(f'rank r(blind, sighted) = {r[0]:.2f}, p = {r[1]:.3f}')

r(blind, sighted) = 0.75, p = 0.000
rank r(blind, sighted) = 0.72, p = 0.000

Predictiveness of different embeddings for the ratings from the original experiment and our first replication¶

Common Crawl¶

In [8]:
m_cc = bambi.Model('rating_z ~ 1'
    + ' + group_eff*frequency_z'
    + ' + group_eff*concreteness_z'
    + ' + group_eff*cosine_cc_z'
    + ' + group_eff*swow_all_z'
    + ' + (1 + frequency_z + concreteness_z + cosine_cc_z + swow_all_z|pp_id)'
    + ' + (1 + group_eff|dimension)'
    + ' + (1 + group_eff|color)',
    df_joint[(df_joint['experiment'] != 'replication_2') & (df_joint['self_vs_other'] == 'self')]
)
r_cc = m_cc.fit(
    init='advi+adapt_diag',
    chains=4,
    draws=1000,
    tune=1000,
    n_init=10000,
    target_accept=.95,
)

Auto-assigning NUTS sampler...
[INFO] Auto-assigning NUTS sampler...
Initializing NUTS using advi+adapt_diag...
[INFO] Initializing NUTS using advi+adapt_diag...
100.00% [10000/10000 00:11<00:00 Average Loss = 16,703] 
Finished [100%]: Average Loss = 16,694
[INFO] Finished [100%]: Average Loss = 16,694
Multiprocess sampling (4 chains in 4 jobs)
[INFO] Multiprocess sampling (4 chains in 4 jobs)
NUTS: [Intercept, group_eff, frequency_z, group_eff:frequency_z, concreteness_z, group_eff:concreteness_z, cosine_cc_z, group_eff:cosine_cc_z, swow_all_z, group_eff:swow_all_z, 1|pp_id_sigma, 1|pp_id_offset, frequency_z|pp_id_sigma, frequency_z|pp_id_offset, concreteness_z|pp_id_sigma, concreteness_z|pp_id_offset, cosine_cc_z|pp_id_sigma, cosine_cc_z|pp_id_offset, swow_all_z|pp_id_sigma, swow_all_z|pp_id_offset, 1|dimension_sigma, 1|dimension_offset, group_eff|dimension_sigma, group_eff|dimension_offset, 1|color_sigma, 1|color_offset, group_eff|color_sigma, group_eff|color_offset, rating_z_sigma]
[INFO] NUTS: [Intercept, group_eff, frequency_z, group_eff:frequency_z, concreteness_z, group_eff:concreteness_z, cosine_cc_z, group_eff:cosine_cc_z, swow_all_z, group_eff:swow_all_z, 1|pp_id_sigma, 1|pp_id_offset, frequency_z|pp_id_sigma, frequency_z|pp_id_offset, concreteness_z|pp_id_sigma, concreteness_z|pp_id_offset, cosine_cc_z|pp_id_sigma, cosine_cc_z|pp_id_offset, swow_all_z|pp_id_sigma, swow_all_z|pp_id_offset, 1|dimension_sigma, 1|dimension_offset, group_eff|dimension_sigma, group_eff|dimension_offset, 1|color_sigma, 1|color_offset, group_eff|color_sigma, group_eff|color_offset, rating_z_sigma]
100.00% [8000/8000 07:35<00:00 Sampling 4 chains, 0 divergences] 
Sampling 4 chains for 1_000 tune and 1_000 draw iterations (4_000 + 4_000 draws total) took 464 seconds.
[INFO] Sampling 4 chains for 1_000 tune and 1_000 draw iterations (4_000 + 4_000 draws total) took 464 seconds.
In [9]:
# compute and plot conditional effect of cosine in sighted and blind groups
trace = deepcopy(r_cc.posterior)
# effect of cosine in blind group in orginal experiment
trace['blind:cosine_cc_z'] = trace['cosine_cc_z'] - trace['group_eff:cosine_cc_z']
    # effect of cosine in sighted group across both experiments
trace['sighted:cosine_cc_z'] = trace['cosine_cc_z'] + trace['group_eff:cosine_cc_z']

varnames = ['blind:cosine_cc_z', 'sighted:cosine_cc_z']
g = az.plot_forest(trace, combined=True, hdi_prob=.95,
                   figsize=[9, .6 + len(varnames) * .3],
                   var_names=varnames)
g[0].axvline(0, color='.8', linewidth=2);
display(az.summary(trace, hdi_prob=.95, var_names=varnames))
mean sd hdi_2.5% hdi_97.5% mcse_mean mcse_sd ess_bulk ess_tail r_hat
blind:cosine_cc_z 0.119 0.066 -0.015 0.24 0.002 0.001 1745.0 2499.0 1.0
sighted:cosine_cc_z 0.388 0.032 0.327 0.45 0.000 0.000 5501.0 3326.0 1.0

OpenSubtitles¶

In [10]:
m_subs = bambi.Model('rating_z ~ 1'
    + ' + group_eff*frequency_z'
    + ' + group_eff*concreteness_z'
    + ' + group_eff*cosine_subs_z'
    + ' + group_eff*swow_all_z'
    + ' + (1 + frequency_z + concreteness_z + cosine_subs_z + swow_all_z|pp_id)'
    + ' + (1 + group_eff|dimension)'
    + ' + (1 + group_eff|color)',
    df_joint[(df_joint['experiment'] != 'replication_2') & (df_joint['self_vs_other'] == 'self')]
)
r_subs = m_subs.fit(
    init='advi+adapt_diag',
    chains=4,
    draws=1000,
    tune=1000,
    n_init=10000,
    target_accept=.95,
)

Auto-assigning NUTS sampler...
[INFO] Auto-assigning NUTS sampler...
Initializing NUTS using advi+adapt_diag...
[INFO] Initializing NUTS using advi+adapt_diag...
100.00% [10000/10000 00:11<00:00 Average Loss = 16,583] 
Finished [100%]: Average Loss = 16,573
[INFO] Finished [100%]: Average Loss = 16,573
Multiprocess sampling (4 chains in 4 jobs)
[INFO] Multiprocess sampling (4 chains in 4 jobs)
NUTS: [Intercept, group_eff, frequency_z, group_eff:frequency_z, concreteness_z, group_eff:concreteness_z, cosine_subs_z, group_eff:cosine_subs_z, swow_all_z, group_eff:swow_all_z, 1|pp_id_sigma, 1|pp_id_offset, frequency_z|pp_id_sigma, frequency_z|pp_id_offset, concreteness_z|pp_id_sigma, concreteness_z|pp_id_offset, cosine_subs_z|pp_id_sigma, cosine_subs_z|pp_id_offset, swow_all_z|pp_id_sigma, swow_all_z|pp_id_offset, 1|dimension_sigma, 1|dimension_offset, group_eff|dimension_sigma, group_eff|dimension_offset, 1|color_sigma, 1|color_offset, group_eff|color_sigma, group_eff|color_offset, rating_z_sigma]
[INFO] NUTS: [Intercept, group_eff, frequency_z, group_eff:frequency_z, concreteness_z, group_eff:concreteness_z, cosine_subs_z, group_eff:cosine_subs_z, swow_all_z, group_eff:swow_all_z, 1|pp_id_sigma, 1|pp_id_offset, frequency_z|pp_id_sigma, frequency_z|pp_id_offset, concreteness_z|pp_id_sigma, concreteness_z|pp_id_offset, cosine_subs_z|pp_id_sigma, cosine_subs_z|pp_id_offset, swow_all_z|pp_id_sigma, swow_all_z|pp_id_offset, 1|dimension_sigma, 1|dimension_offset, group_eff|dimension_sigma, group_eff|dimension_offset, 1|color_sigma, 1|color_offset, group_eff|color_sigma, group_eff|color_offset, rating_z_sigma]
100.00% [8000/8000 07:41<00:00 Sampling 4 chains, 0 divergences] 
Sampling 4 chains for 1_000 tune and 1_000 draw iterations (4_000 + 4_000 draws total) took 470 seconds.
[INFO] Sampling 4 chains for 1_000 tune and 1_000 draw iterations (4_000 + 4_000 draws total) took 470 seconds.
In [11]:
# compute and plot conditional effect of cosine in sighted and blind groups
trace_subs = deepcopy(r_subs.posterior)
# effect of cosine in blind group in orginal experiment
trace['blind:cosine_subs_z'] = trace_subs['cosine_subs_z'] - trace_subs['group_eff:cosine_subs_z']
# effect of cosine in sighted group across both experiments
trace['sighted:cosine_subs_z'] = trace_subs['cosine_subs_z'] + trace_subs['group_eff:cosine_subs_z']

varnames = [
    'blind:cosine_cc_z', 'sighted:cosine_cc_z',
    'blind:cosine_subs_z', 'sighted:cosine_subs_z',
]
g = az.plot_forest(trace, combined=True, hdi_prob=.95,
                   figsize=[9, .6 + len(varnames) * .3],
                   var_names=varnames)
g[0].axvline(0, color='.8', linewidth=2);
display(az.summary(trace, hdi_prob=.95, var_names=varnames))
mean sd hdi_2.5% hdi_97.5% mcse_mean mcse_sd ess_bulk ess_tail r_hat
blind:cosine_cc_z 0.119 0.066 -0.015 0.240 0.002 0.001 1745.0 2499.0 1.0
sighted:cosine_cc_z 0.388 0.032 0.327 0.450 0.000 0.000 5501.0 3326.0 1.0
blind:cosine_subs_z 0.225 0.047 0.134 0.315 0.001 0.001 2846.0 3226.0 1.0
sighted:cosine_subs_z 0.446 0.024 0.398 0.493 0.000 0.000 6598.0 3293.0 1.0

COCA-fiction¶

In [5]:
m_fic = bambi.Model('rating_z ~ 1'
    + ' + group_eff*frequency_z'
    + ' + group_eff*concreteness_z'
    + ' + group_eff*cosine_fic_z'
    + ' + group_eff*swow_all_z'
    + ' + (1 + frequency_z + concreteness_z + cosine_fic_z + swow_all_z|pp_id)'
    + ' + (1 + group_eff|dimension)'
    + ' + (1 + group_eff|color)',
    df_joint[(df_joint['experiment'] != 'replication_2') & (df_joint['self_vs_other'] == 'self')]
)
r_fic = m_fic.fit(
    init='advi+adapt_diag',
    chains=4,
    draws=1000,
    tune=1000,
    n_init=10000,
    target_accept=.95,
)

Auto-assigning NUTS sampler...
[INFO] Auto-assigning NUTS sampler...
Initializing NUTS using advi+adapt_diag...
[INFO] Initializing NUTS using advi+adapt_diag...
100.00% [10000/10000 00:13<00:00 Average Loss = 16,641] 
Finished [100%]: Average Loss = 16,632
[INFO] Finished [100%]: Average Loss = 16,632
Multiprocess sampling (4 chains in 4 jobs)
[INFO] Multiprocess sampling (4 chains in 4 jobs)
NUTS: [Intercept, group_eff, frequency_z, group_eff:frequency_z, concreteness_z, group_eff:concreteness_z, cosine_fic_z, group_eff:cosine_fic_z, swow_all_z, group_eff:swow_all_z, 1|pp_id_sigma, 1|pp_id_offset, frequency_z|pp_id_sigma, frequency_z|pp_id_offset, concreteness_z|pp_id_sigma, concreteness_z|pp_id_offset, cosine_fic_z|pp_id_sigma, cosine_fic_z|pp_id_offset, swow_all_z|pp_id_sigma, swow_all_z|pp_id_offset, 1|dimension_sigma, 1|dimension_offset, group_eff|dimension_sigma, group_eff|dimension_offset, 1|color_sigma, 1|color_offset, group_eff|color_sigma, group_eff|color_offset, rating_z_sigma]
[INFO] NUTS: [Intercept, group_eff, frequency_z, group_eff:frequency_z, concreteness_z, group_eff:concreteness_z, cosine_fic_z, group_eff:cosine_fic_z, swow_all_z, group_eff:swow_all_z, 1|pp_id_sigma, 1|pp_id_offset, frequency_z|pp_id_sigma, frequency_z|pp_id_offset, concreteness_z|pp_id_sigma, concreteness_z|pp_id_offset, cosine_fic_z|pp_id_sigma, cosine_fic_z|pp_id_offset, swow_all_z|pp_id_sigma, swow_all_z|pp_id_offset, 1|dimension_sigma, 1|dimension_offset, group_eff|dimension_sigma, group_eff|dimension_offset, 1|color_sigma, 1|color_offset, group_eff|color_sigma, group_eff|color_offset, rating_z_sigma]
100.00% [8000/8000 07:36<00:00 Sampling 4 chains, 0 divergences] 
Sampling 4 chains for 1_000 tune and 1_000 draw iterations (4_000 + 4_000 draws total) took 466 seconds.
[INFO] Sampling 4 chains for 1_000 tune and 1_000 draw iterations (4_000 + 4_000 draws total) took 466 seconds.
In [17]:
# compute and plot conditional effect of cosine in sighted and blind groups
trace_fic = deepcopy(r_fic.posterior)
# effect of cosine in blind group in orginal experiment
trace['blind:cosine_fic_z'] = trace_fic['cosine_fic_z'] - trace_fic['group_eff:cosine_fic_z']
# effect of cosine in sighted group across both experiments
trace['sighted:cosine_fic_z'] = trace_fic['cosine_fic_z'] + trace_fic['group_eff:cosine_fic_z']

varnames = [
    'blind:cosine_cc_z', 'sighted:cosine_cc_z',
    'blind:cosine_subs_z', 'sighted:cosine_subs_z',
    'blind:cosine_fic_z', 'sighted:cosine_fic_z',
]
g = az.plot_forest(trace, combined=True, hdi_prob=.95,
                   figsize=[9, .6 + len(varnames) * .3],
                   var_names=varnames)
g[0].axvline(0, color='.8', linewidth=2);
display(az.summary(trace, hdi_prob=.95, var_names=varnames))
mean sd hdi_2.5% hdi_97.5% mcse_mean mcse_sd ess_bulk ess_tail r_hat
blind:cosine_cc_z 0.119 0.066 -0.015 0.240 0.002 0.001 1745.0 2499.0 1.0
sighted:cosine_cc_z 0.388 0.032 0.327 0.450 0.000 0.000 5501.0 3326.0 1.0
blind:cosine_subs_z 0.225 0.047 0.134 0.315 0.001 0.001 2846.0 3226.0 1.0
sighted:cosine_subs_z 0.446 0.024 0.398 0.493 0.000 0.000 6598.0 3293.0 1.0
blind:cosine_fic_z 0.340 0.053 0.241 0.445 0.001 0.001 1912.0 2704.0 1.0
sighted:cosine_fic_z 0.495 0.027 0.444 0.548 0.000 0.000 5625.0 2970.0 1.0

COCA-spoken¶

In [6]:
m_spok = bambi.Model('rating_z ~ 1'
    + ' + group_eff*frequency_z'
    + ' + group_eff*concreteness_z'
    + ' + group_eff*cosine_spok_z'
    + ' + group_eff*swow_all_z'
    + ' + (1 + frequency_z + concreteness_z + cosine_spok_z + swow_all_z|pp_id)'
    + ' + (1 + group_eff|dimension)'
    + ' + (1 + group_eff|color)',
    df_joint[(df_joint['experiment'] != 'replication_2') & (df_joint['self_vs_other'] == 'self')]
)
r_spok = m_spok.fit(
    init='advi+adapt_diag',
    chains=4,
    draws=1000,
    tune=1000,
    n_init=10000,
    target_accept=.95,
)

Auto-assigning NUTS sampler...
[INFO] Auto-assigning NUTS sampler...
Initializing NUTS using advi+adapt_diag...
[INFO] Initializing NUTS using advi+adapt_diag...
100.00% [10000/10000 00:11<00:00 Average Loss = 16,722] 
Finished [100%]: Average Loss = 16,708
[INFO] Finished [100%]: Average Loss = 16,708
Multiprocess sampling (4 chains in 4 jobs)
[INFO] Multiprocess sampling (4 chains in 4 jobs)
NUTS: [Intercept, group_eff, frequency_z, group_eff:frequency_z, concreteness_z, group_eff:concreteness_z, cosine_spok_z, group_eff:cosine_spok_z, swow_all_z, group_eff:swow_all_z, 1|pp_id_sigma, 1|pp_id_offset, frequency_z|pp_id_sigma, frequency_z|pp_id_offset, concreteness_z|pp_id_sigma, concreteness_z|pp_id_offset, cosine_spok_z|pp_id_sigma, cosine_spok_z|pp_id_offset, swow_all_z|pp_id_sigma, swow_all_z|pp_id_offset, 1|dimension_sigma, 1|dimension_offset, group_eff|dimension_sigma, group_eff|dimension_offset, 1|color_sigma, 1|color_offset, group_eff|color_sigma, group_eff|color_offset, rating_z_sigma]
[INFO] NUTS: [Intercept, group_eff, frequency_z, group_eff:frequency_z, concreteness_z, group_eff:concreteness_z, cosine_spok_z, group_eff:cosine_spok_z, swow_all_z, group_eff:swow_all_z, 1|pp_id_sigma, 1|pp_id_offset, frequency_z|pp_id_sigma, frequency_z|pp_id_offset, concreteness_z|pp_id_sigma, concreteness_z|pp_id_offset, cosine_spok_z|pp_id_sigma, cosine_spok_z|pp_id_offset, swow_all_z|pp_id_sigma, swow_all_z|pp_id_offset, 1|dimension_sigma, 1|dimension_offset, group_eff|dimension_sigma, group_eff|dimension_offset, 1|color_sigma, 1|color_offset, group_eff|color_sigma, group_eff|color_offset, rating_z_sigma]
100.00% [8000/8000 07:20<00:00 Sampling 4 chains, 0 divergences] 
Sampling 4 chains for 1_000 tune and 1_000 draw iterations (4_000 + 4_000 draws total) took 449 seconds.
[INFO] Sampling 4 chains for 1_000 tune and 1_000 draw iterations (4_000 + 4_000 draws total) took 449 seconds.
In [19]:
# compute and plot conditional effect of cosine in sighted and blind groups
trace_spok = deepcopy(r_spok.posterior)
# effect of cosine in blind group in orginal experiment
trace['blind:cosine_spok_z'] = trace_spok['cosine_spok_z'] - trace_spok['group_eff:cosine_spok_z']
# effect of cosine in sighted group across both experiments
trace['sighted:cosine_spok_z'] = trace_spok['cosine_spok_z'] + trace_spok['group_eff:cosine_spok_z']

varnames = [
    'blind:cosine_cc_z', 'sighted:cosine_cc_z',
    'blind:cosine_subs_z', 'sighted:cosine_subs_z',
    'blind:cosine_fic_z', 'sighted:cosine_fic_z',
    'blind:cosine_spok_z', 'sighted:cosine_spok_z',
]
g = az.plot_forest(trace, combined=True, hdi_prob=.95,
                   figsize=[9, .6 + len(varnames) * .3],
                   var_names=varnames)
g[0].axvline(0, color='.8', linewidth=2);
display(az.summary(trace, hdi_prob=.95, var_names=varnames))
mean sd hdi_2.5% hdi_97.5% mcse_mean mcse_sd ess_bulk ess_tail r_hat
blind:cosine_cc_z 0.119 0.066 -0.015 0.240 0.002 0.001 1745.0 2499.0 1.0
sighted:cosine_cc_z 0.388 0.032 0.327 0.450 0.000 0.000 5501.0 3326.0 1.0
blind:cosine_subs_z 0.225 0.047 0.134 0.315 0.001 0.001 2846.0 3226.0 1.0
sighted:cosine_subs_z 0.446 0.024 0.398 0.493 0.000 0.000 6598.0 3293.0 1.0
blind:cosine_fic_z 0.340 0.053 0.241 0.445 0.001 0.001 1912.0 2704.0 1.0
sighted:cosine_fic_z 0.495 0.027 0.444 0.548 0.000 0.000 5625.0 2970.0 1.0
blind:cosine_spok_z 0.055 0.041 -0.028 0.135 0.001 0.001 1878.0 2455.0 1.0
sighted:cosine_spok_z 0.145 0.024 0.096 0.191 0.000 0.000 5146.0 3389.0 1.0

Combined conditional effects plot (effects of each different embedding projection in the blind and sighted groups)¶

In [20]:
varnames = [
    'sighted:cosine_cc_z',
    'sighted:cosine_subs_z',
    'sighted:cosine_fic_z',
    'sighted:cosine_spok_z',
    'blind:cosine_cc_z',
    'blind:cosine_subs_z',
    'blind:cosine_fic_z',
    'blind:cosine_spok_z',
]
trace_df = az.summary(trace, hdi_prob=.95, var_names=varnames).reset_index()
sns.set(palette='colorblind', style='whitegrid')
trace_df['group'] = trace_df['index'].apply(lambda x: x.split(':')[0])
trace_df['predictor'] = trace_df['index'].apply(lambda x: x.split(':')[1])
g, ax = plt.subplots(figsize=[3.2, 3.2])
ax.axvline(0, color='.8', linewidth=2)
varnames = [
    'sighted:cosine_cc_z',
    'sighted:cosine_subs_z',
    'sighted:cosine_fic_z',
    'sighted:cosine_spok_z',
    'blind:cosine_cc_z',
    'blind:cosine_subs_z',
    'blind:cosine_fic_z',
    'blind:cosine_spok_z',
]
order = [
    'cosine_cc_z',
    'cosine_subs_z',
    'cosine_fic_z',
    'cosine_spok_z',
]
sns.pointplot(data=trace_df, hue='group', x='mean', y='predictor', join=False, ax=ax, order=order,
              palette=sns.color_palette()[0:], dodge=.2)
hdi_df = trace_df.melt(id_vars=['predictor', 'group'], value_vars=['hdi_2.5%', 'hdi_97.5%'])
sns.pointplot(data=hdi_df, hue='group', markers='', x='value', order=order,
              y='predictor', ax=ax, join=False,
              palette=sns.color_palette()[0:], dodge=.2)
ax.set(xlim=[-.1, .6], xticks=[-.1, 0, .1, .2, .3, .4, .5, .6],
       ylabel='', xlabel='effect size (standardized coefficient)',
       title='95% CIs for embedding projections\nfrom various corpora')
ax.set(yticklabels=[
    'Common Crawl projection',
    'OpenSubtitles projection',
    'COCA-fiction projection',
    'COCA-spoken projection',
])
labels = ax.get_legend_handles_labels()
ax.legend(handles=labels[0][:2], labels=labels[1][:2],
          bbox_to_anchor=(1, 1), loc=2, borderaxespad=0, frameon=False)
plt.savefig('figures/corpus_variation_forest.pdf', bbox_inches='tight')

/opt/homebrew/lib/python3.9/site-packages/seaborn/categorical.py:1727: UserWarning: You passed a edgecolor/edgecolors ((0.00392156862745098, 0.45098039215686275, 0.6980392156862745)) for an unfilled marker ('').  Matplotlib is ignoring the edgecolor in favor of the facecolor.  This behavior may change in the future.
  ax.scatter(x, y, label=hue_level,
/opt/homebrew/lib/python3.9/site-packages/seaborn/categorical.py:1727: UserWarning: You passed a edgecolor/edgecolors ((0.8705882352941177, 0.5607843137254902, 0.0196078431372549)) for an unfilled marker ('').  Matplotlib is ignoring the edgecolor in favor of the facecolor.  This behavior may change in the future.
  ax.scatter(x, y, label=hue_level,

Comparing models on leave-one-out validation score¶

Instead of using conditional effect sizes from the models as a comparison, we can also compare the models on their out-of-sample predictive accuracy. Essentially we're looking to see which predictor, if included, makes the model worst.
Using prediction/cross-validation for model comparison has distinct advantages in some situations, but in our case (when we're not varying model complexity or structure, but only replacing a single predictor) the results will probably match our inferences from the conditional effect size plots.

In [21]:
display(az.compare({
    'COCA fiction': r_fic,
    'COCA spoken': r_spok,
    'Common Crawl': r_cc,
    'OpenSubtitles': r_subs,
}).round(2))

/opt/homebrew/lib/python3.9/site-packages/arviz/stats/stats.py:802: UserWarning: Estimated shape parameter of Pareto distribution is greater than 0.7 for one or more samples. You should consider using a more robust model, this is because importance sampling is less likely to work well if the marginal posterior and LOO posterior are very different. This is more likely to happen with a non-robust model and highly influential observations.
  warnings.warn(
/opt/homebrew/lib/python3.9/site-packages/arviz/stats/stats.py:802: UserWarning: Estimated shape parameter of Pareto distribution is greater than 0.7 for one or more samples. You should consider using a more robust model, this is because importance sampling is less likely to work well if the marginal posterior and LOO posterior are very different. This is more likely to happen with a non-robust model and highly influential observations.
  warnings.warn(
/opt/homebrew/lib/python3.9/site-packages/arviz/stats/stats.py:802: UserWarning: Estimated shape parameter of Pareto distribution is greater than 0.7 for one or more samples. You should consider using a more robust model, this is because importance sampling is less likely to work well if the marginal posterior and LOO posterior are very different. This is more likely to happen with a non-robust model and highly influential observations.
  warnings.warn(
/opt/homebrew/lib/python3.9/site-packages/arviz/stats/stats.py:802: UserWarning: Estimated shape parameter of Pareto distribution is greater than 0.7 for one or more samples. You should consider using a more robust model, this is because importance sampling is less likely to work well if the marginal posterior and LOO posterior are very different. This is more likely to happen with a non-robust model and highly influential observations.
  warnings.warn(
rank elpd_loo p_loo elpd_diff weight se dse warning scale
COCA fiction 0 -13497.35 157.23 0.00 0.52 64.91 0.00 True log
OpenSubtitles 1 -13505.92 162.67 8.57 0.48 65.04 19.67 True log
Common Crawl 2 -13608.02 160.27 110.68 0.00 64.65 19.18 True log
COCA spoken 3 -13687.66 155.01 190.32 0.00 64.16 19.76 True log

As expected, the model with COCA-fiction performs best, although the subtitle model is a close second. These models have in common that they contain many well-formed sentences of a decent length, which contain references to color in both literal and metaphorical sense.

The comparison algorithm does warn that the shape parameter of the pareto distribution is too large for some observations. This is common with hierarchical models and shouldn't be a problem if it concerns only a few observations. We can check this by inspecting the LOO-statistics more closely.

In [22]:
az.loo(r_fic)

/opt/homebrew/lib/python3.9/site-packages/arviz/stats/stats.py:802: UserWarning: Estimated shape parameter of Pareto distribution is greater than 0.7 for one or more samples. You should consider using a more robust model, this is because importance sampling is less likely to work well if the marginal posterior and LOO posterior are very different. This is more likely to happen with a non-robust model and highly influential observations.
  warnings.warn(
Out[22]:
Computed from 4000 posterior samples and 9417 observations log-likelihood matrix.

         Estimate       SE
elpd_loo -13497.35    64.91
p_loo      157.23        -

There has been a warning during the calculation. Please check the results.
------

Pareto k diagnostic values:
                         Count   Pct.
(-Inf, 0.5]   (good)     9408   99.9%
 (0.5, 0.7]   (ok)          7    0.1%
   (0.7, 1]   (bad)         2    0.0%
   (1, Inf)   (very bad)    0    0.0%

It appears that for the COCA-fiction model, for instance, there are only 3 observations in the "bad" range. This is not overly problematic given the nearly 10,000 observations that are in the correct range.

Visual model diagnostics¶

R-hat and other Markov chain diagnostics looked good, but it's still worth doing a few quick visual model checks.

Quick look at the trace plots¶

We want all the Markov chains for a single variable to look lined up in the density plot on the left, but nice and fuzzy (i.e. not autocorrelated) on the right.

In [27]:
g = az.plot_trace(r_cc, var_names=['Intercept', 'group_eff', 'cosine_cc_z', 'group_eff:cosine_cc_z'])
In [28]:
g = az.plot_trace(r_subs, var_names=['Intercept', 'group_eff', 'cosine_subs_z', 'group_eff:cosine_subs_z'])
In [29]:
g = az.plot_trace(r_fic, var_names=['Intercept', 'group_eff', 'cosine_fic_z', 'group_eff:cosine_fic_z'])
In [30]:
g = az.plot_trace(r_spok, var_names=['Intercept', 'group_eff', 'cosine_spok_z', 'group_eff:cosine_spok_z'])

Cumulative posterior predictive check¶

Check if the model's predictions line up with the predictions in our observed data (i.e. there are no weird biases etc.)

In [14]:
m_cc.predict(r_cc, kind='pps')
g = az.plot_ppc(r_cc, kind='cumulative')
In [15]:
m_subs.predict(r_subs, kind='pps')
g = az.plot_ppc(r_subs, kind='cumulative')
In [7]:
m_fic.predict(r_fic, kind='pps')
g = az.plot_ppc(r_fic, kind='cumulative')
In [8]:
m_spok.predict(r_spok, kind='pps')
g = az.plot_ppc(r_spok, kind='cumulative')

Convert notebook to html¶

In [27]:
convert_notebook('experiment_2')

Jupyter notebook experiment_2 converted successfully.

In [ ]: